frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 14:53:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Config] add "qwen" as a native eagle3 target supported model (#22333)

Browse Source 
Signed-off-by: lechen <lecself@163.com>
Signed-off-by: LeChen <lecself@163.com>
This commit is contained in:
Le Chen
2025-08-10 11:21:05 +08:00
committed by GitHub
parent 0c5254b82a
commit 3d7363e61c
5 changed files with 30 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/models/registry.py
View File

@ -525,6 +525,10 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
trust_remote_code=True,
speculative_model="yuhuili/EAGLE3-LLaMA3.1-Instruct-8B",
tokenizer="meta-llama/Llama-3.1-8B-Instruct"),
"LlamaForCausalLMEagle3": _HfExamplesInfo("AngelSlim/Qwen3-8B_eagle3", # noqa: E501
trust_remote_code=True,
speculative_model="AngelSlim/Qwen3-8B_eagle3",
tokenizer="Qwen/Qwen3-8B"),
"EagleLlama4ForCausalLM": _HfExamplesInfo(
"morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct",
trust_remote_code=True,

39
tests/v1/e2e/test_spec_decode.py
View File

@ -125,24 +125,27 @@ def test_ngram_correctness(
cleanup_dist_env_and_memory()
@pytest.mark.parametrize(
["model_setup", "mm_enabled"], [
(("eagle", "meta-llama/Llama-3.1-8B-Instruct",
"yuhuili/EAGLE-LLaMA3.1-Instruct-8B", 1), False),
(("eagle3", "meta-llama/Llama-3.1-8B-Instruct",
"yuhuili/EAGLE3-LLaMA3.1-Instruct-8B", 1), False),
pytest.param(
("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
"morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
False,
marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
pytest.param(
("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
"morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
True,
marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
],
ids=["llama3_eagle", "llama3_eagle3", "llama4_eagle", "llama4_eagle_mm"])
@pytest.mark.parametrize(["model_setup", "mm_enabled"], [
(("eagle3", "Qwen/Qwen3-8B", "AngelSlim/Qwen3-8B_eagle3", 1), False),
(("eagle", "meta-llama/Llama-3.1-8B-Instruct",
"yuhuili/EAGLE-LLaMA3.1-Instruct-8B", 1), False),
(("eagle3", "meta-llama/Llama-3.1-8B-Instruct",
"yuhuili/EAGLE3-LLaMA3.1-Instruct-8B", 1), False),
pytest.param(
("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
"morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
False,
marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
pytest.param(
("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
"morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
True,
marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
],
ids=[
"qwen3_eagle3", "llama3_eagle", "llama3_eagle3",
"llama4_eagle", "llama4_eagle_mm"
])
@pytest.mark.parametrize("attn_backend",
get_attn_backend_list_based_on_platform())
def test_eagle_correctness(

8
vllm/config/__init__.py
View File

@ -2852,13 +2852,7 @@ class SpeculativeConfig:
"speculative decoding is > 1, but got "
f"{self.disable_by_batch_size=}")
from vllm.transformers_utils.configs import SpeculatorsConfig
eagle3_target_supported = ["llama"]
if self.draft_model_config and isinstance(
self.draft_model_config.hf_config, SpeculatorsConfig):
eagle3_target_supported.append("qwen")
eagle3_target_supported = ["llama", "qwen"]
if self.method == "eagle3" and self.target_model_config and not any(
supported_model in
self.target_model_config.hf_text_config.model_type

1
vllm/model_executor/models/registry.py
View File

@ -259,6 +259,7 @@ _SPECULATIVE_DECODING_MODELS = {
"EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
"EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
"Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
"LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
"DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
"Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
"MedusaModel": ("medusa", "Medusa"),

5
vllm/transformers_utils/configs/eagle.py
View File

@ -45,6 +45,7 @@ class EAGLEConfig(PretrainedConfig):
# Eagle model name should follow naming convention of
# LlamaForCausalLM -> EagleLlamaForCausalLM
# LlamaForCausalLM -> Eagle3LlamaForCausalLM / LlamaForCausalLMEagle3
if method == "eagle":
assert self.model is not None, \
"model should not be None when method is eagle"
@ -56,8 +57,8 @@ class EAGLEConfig(PretrainedConfig):
assert self.model is not None, \
"model should not be None when method is eagle3"
kwargs["architectures"] = [
f"Eagle3{arch}" if not arch.startswith("Eagle3") \
else arch for arch in self.model.architectures
arch if arch.startswith("Eagle3") or arch.endswith("Eagle3")
else f"Eagle3{arch}" for arch in self.model.architectures
]
else:
raise ValueError(f"Invalid method {method}. \